Imbalance-Aware Small-Data Machine Learning for Dimensionality Prediction in Hybrid Metal Halides

In this work, we present an imbalance-aware, small-data machine-learning workflow to predict the structural dimensionality (0D–3D) of hybrid metal halides, one of the key factors governing exciton binding, charge transport, and environmental stability. Our model is first trained on a highly-imbalanced hybrid metal halide dataset from the HybriD3 database ( ≈ 67% 2D). Chemically informed descriptors capture steric and connectivity effects, while interaction-based descriptors further account for nonlinear relationships. We address data imbalance with synthetic oversampling and targeted feature engineering, followed by the development of stacked ensemble classifiers to improve accuracy. This approach enhances minority-class accuracy while maintaining overall performance. Five-fold cross-validation yields a mean F1 score of 0.964 with low variance, demonstrating strong generalization. The developed framework is interpretable, efficient, and applicable to other imbalanced, small-scale materials datasets.